A signal prediction simulation tool for ultrasonic testing (UT) has been developed by combining the finite integration technique (FIT) with an image-based modeling approach. A realistic numerical model of a target can be made directly from digital images such as X-ray CT and three dimensional point cloud data. The image-based FIT for UT is accelerated using graphics processing units (GPUs) with Compute Unified Device Architecture (CUDA) Fortran. The methodology is described, and checked by numerical experiments using multiple GPU boards. The calculation speed can be dramatically improved compared to that obtained by running the same simulation on a conventional CPU. As an application of the image-based FIT, we propose a simulation-aided flaw reconstruction method for UT by means of the time reversal approach. Scattered waves from a flaw are recorded using an array transducer, and the time-reversed waves are re-emitted to numerical model in the FIT simulation. The re-emitted waves propagate through the heterogeneous media and focus on the flaw. The shape of the flaw can be visually estimated from the focal point of the ultrasonic wave in the simulation.